Language is full of patterns and regularities. In the last decades, there has been great interest in the role of statistical learning in language acquisition. Statistical learning is a cognitive ability that underlies the implicit discovery of statistical patterns and sequences in sensory input (Siegelman et al., 2017) and has been hypothesized to contribute to different areas of language acquisition (for a review, see Romberg & Saffran, 2010). One of the first demonstrations of statistical learning was the seminal study of Saffran, Aslin and Newport (1996). As word boundaries are not (consistently) marked by pauses or other prosodic cues in natural speech (Cole, 1980), the authors aimed to investigate whether statistical learning plays a role in learning to recognize separate words in a stream of speech sounds. Eight-month-old infants with English-speaking parents were exposed to a two-minute synthesized stream of uninterrupted syllables. The speech stream consisted of four pseudo-words (bidaku, padoti, golabu and tupiro) that were repeated in a random order. The authors wanted to test whether infants were able to recognize these pseudo-words after exposure to the stream, despite the absence of any prosodic cues for word boundaries. Results from a head-turn preference procedure administered after familiarization show that infants listen longer to “part-words” that span word boundaries, such as ku-pado, than to target words. This novelty preference indicates that infants learn to recognize target words and are thus sensitive to the statistical probabilities of the input: the transitional probabilities (TP s) between syllables. For example, the probability that da followed bi in the stream was 1.0, while the probability that pa followed ku was only 0.333. As there were no pauses or other prosodic cues for word boundaries,¹ infants’ learning could only have happened based on these TP values.

The degree of learning in statistical learning tasks is usually inferred from participants’ performance on an “offline” task which they undergo after the familiarization phase, during which they have to choose between target words and foils. However, performance on such tasks could be strongly influenced by cognitive processes other than statistical learning, such as encoding and memory capacities, meta-linguistic skills and decision-making biases (Siegelman et al., 2017). Specifically for children, meta-linguistic questions such as “which word sounds better?” are difficult to process and answer, which could lead to underestimation of their (implicit) knowledge. Importantly, while statistical learning is a continuous process, offline measures provide information about behavior at only a single point in time. Online methods, on the other hand, can provide more insight into the trajectory of statistical learning by measuring learning throughout the familiarization phase. It has thus been argued that in future statistical learning studies, especially those focusing on children, it is important to develop sensitive online measures of statistical learning (Lammertink et al., 2019; Siegelman et al., 2017).

Based on the idea that reaction time reflects processing time, stimulus detection tasks have been used as online measures of sentence processing (e.g. Fodor & Bever, 1965; Foss & Lynch, 1969; Cohen & Mehler, 1996), and this paradigm has also been applied to word segmentation tasks. Gómez et al. (2011) added click sounds to the speech stream in their word segmentation task. Italian-speaking adults were asked to listen to a stream of speech sounds, consisting of four pseudo-words (pabuda, gifoto, venola and minaro) for four minutes and to push a button as fast as possible when they heard a click sound. Crucially, the clicks occurred either between two pseudo-words or within a pseudo-word (compare pabuda!gifoto to pa!buda, where ! indicates a click). The authors hypothesized that participants who had learned to recognize words in the stream should have stronger expectations for the next syllable when hearing the first syllable of a word compared to when they hear the final syllable of a word. This, in turn, should lead to a larger surprise effect (and thus a slower reaction time) when detecting clicks occurring within words compared to clicks between words. Results showed that after two minutes of exposure, people are indeed slower when detecting clicks within words as opposed to clicks between words, indicating sensitivity to word boundaries that develops over time.

Franco et al. (2015) aimed to replicate these results and tested French-speaking adults on a similar task. As opposed to Gómez et al. (2011), the researchers did not find evidence for a difference in response times to clicks between words and clicks within words. Ten out of 28 participants showed the expected pattern while the other 18 showed the opposite pattern. In their second experiment, Franco et al. (2015) compared performance on two versions of the task: a “passive” word segmentation task with clicks to which participants did not have to respond (“passive-click”); and a word segmentation task without any clicks (“no-click”). They found that performance on the offline test phase of the passive-click version was significantly lower than performance on the no-click condition, indicating that the statistical learning process might have suffered from the addition of clicks to the stream. Hearing the clicks might have diverted attention from the syllable structure in the input, as participants might have focused more on detecting the clicks than on the artificial language. Another possibility is that the clicks might have given participants false cues to word boundaries, as the clicks were the only “prosodic” elements in the speech stream. The click detection paradigm has the potential to reveal the word segmentation process minute by minute,² but the finding of mixed results might indicate that an adaptation of the paradigm is called for.

Our aim was to find a method for measuring word segmentation online that would be suitable for adults as well as for children. As Gómez et al. (2011) and Franco et al. (2015) found mixed results, we decided to adapt the click detection by extending the familiarization phase to eight minutes to facilitate learning. The first and final two minutes contained only a few click sounds and were added to provide the participants with more “clean” input (without potential distraction from clicks) to facilitate learning of word boundaries. Based on previous studies, we hypothesized that participants could use statistical information to segment words from uninterrupted speech and that our adaptations to the task would result in a learning effect: slower reaction times for clicks within words compared to clicks between words. We constructed an artificial language based on the study of Haebig et al. (2017), as they tested a similar participant group as we intend to test for our future studies (school-aged typically developing children and children with developmental language disorder; DLD). We conducted three separate experiments. In our first experiment we tested online word segmentation using the click detection task. As we did not find evidence for learning on either the click detection nor the offline task, we conducted a second experiment in which we removed the click sounds to test whether participants (adults and children) would show learning on the offline task. Finally, as we did not find evidence for offline learning in Experiment 2, we conducted Experiment 3 in which we used non-words (TP = 0) as foils instead of part-words (TP = 0.333), to test whether adults would learn to distinguish words from non-words.

In the present set of studies, we aimed to find a method for measuring word segmentation online that was suitable for testing school-aged children, but we encountered some unexpected outcomes. Firstly, the click detection task does not seem to be a reliable measure of online word segmentation. There was a large amount of individual variation in the data, and test version may have influenced the results. We cannot state that responding to click sounds in a stream of interrupted syllables consistently revealed an effect of word boundary knowledge. The fact that our artificial language consisted of bisyllabic words means that the click sounds occurring within words always divided the words in two single syllables. This is not the case for a language containing trisyllabic words (Franco et al., 2015; Gómez et al., 2011), as within-word clicks in that case still leave two syllables of that word uninterrupted. This repetition of high transitional probabilities (between syllable 1 and 2 and between syllable 2 and 3) might counter the effect of clicks being perceived as cues for word boundaries. Future studies might adapt our paradigm (adding blocks with only a few click sounds to the familiarization phase) using trisyllabic words to test whether the click detection task would reveal online learning then. A general difficulty with the click detection paradigm might be the so-called “auditory streaming effect” or “auditory stream segregation” (Micheyl et al., 2010; Van Noorden, 1975): participants might perceive the syllables and the click sounds as two separate sound streams. If this is the case (for some listeners), this might be the reason that the position of the click sounds does not influence their reaction time to them. Participants reported that they found it very hard to pinpoint in which specific syllable a click occurred. Moreover, the addition of click sounds to the stream could have distracted the listeners’ attention away from the to-be-learned word boundaries, as suggested by Franco et al. (2015) and Toro et al. (2005). The below-chance performance on the offline test phase in our first experiment seemed to point in that direction. We wanted to investigate this by conducting Experiment 2, in which we tested the same word segmentation task without the addition of the click detection task.

Contrary to our expectations, we did not find evidence for or against sensitivity to words boundaries in adults or children in Experiment 2. This was also the case for our third experiment, in which the foils in the test phase were not part-words (TP = 0.333) but non-words (TP = 0). There are multiple factors that could have influenced our results. First, it could be the case that two-syllable words are somehow too short to “trigger” a statistical learning mechanism, although Graf Estes and Lew-Williams (2015) and Haebig et al. (2017), who also used bisyllabic words, did find a learning effect. However, differently from Haebig et al. (2017), our participants listened to the language for 8 minutes instead of 4.75 minutes. It is possible that the high amount of exposure to the syllables had given the participants the impression that the language consisted of monosyllabic words instead of bisyllabic words, resulting in less sensitivity to syllable combinations. Second, we used natural modified speech instead of synthesized speech. In a meta-analysis, Black and Bergmann (2017) found that infants’ word segmentation ability was stronger in experiments that use synthesized speech. Natural speech contains more information than synthesized speech which could make the processing of the stream and consequently learning of word boundaries more difficult. Regarding the test phase, almost all participants stated that they found it difficult and often also reported that the test became more difficult as it progressed. This might be due to the repetition of the targets and foils in the test phase, which could overwrite the (weak) representations that might have been built during the familiarization phase (Siegelman et al., 2017).

Influence of prior linguistic knowledge could also have played a role (Finn & Kam, 2008; Siegelman et al., 2018; Van Hedger et al., 2020). Participants, especially adults or older children, who are subjected to an artificial language in a word segmentation task are not blank slates but already have linguistic knowledge and thus expectations about sounds and sound combinations. This knowledge might influence the learning process. As this influence is hard to predict correctly, the particular words that are chosen in an experiment might impact participants’ performance. This is for example illustrated by findings of Erickson et al. (2016), who tested participants on two word segmentation tasks with different sets of words. Performance on one task did not reliably predict performance on the other task. Siegelman et al. (2018) suggest that the influence of prior linguistic knowledge (or “entrenchment effects”) plays a very important factor in the large differences in effect sizes and reliability that is found between statistical learning studies. The influence of prior linguistic knowledge might in some cases be stronger than the influence of the statistical properties of the input that participants are briefly subjected to. Entrenchment effects could have led to the null results and the unexpected version effects in Experiment 1 and 3 of the current study and possibly more studies that have ended up in drawers. Future research should investigate this phenomenon in depth.

In sum, measuring word segmentation ability reliably might be more sensitive to methodological choices than assumed (see also Black and Bergmann, 2017). We would like to emphasize that studies that fail to find a significant effect may not be published (the “file drawer effect”). Access to null results is essential for reliable meta-analyses, which are an important source of empirical evidence. Therefore, it is important to report the results of our current study. Future research should systematically investigate what constraints the word segmentation ability and how (online) learning can be detected reliably (Siegelman et al., 2017). For example, Lukács et al. (2019) tested out an a method of measuring word segmentation online with a syllable detection task, and Kidd et al. (2020) used a serial recall task to measure offline learning more reliably.

Many thanks to our lab technician Dirk-Jan Vet for his help with the development of Praat and E-Prime scripts and providing the test equipment. Moreover, we want to thank Jeremy Bos, Lia Morrissey, Sonja Helder and Maayke Sterk for their help in recruiting and testing participants.